TokenPad

Prompting

System prompt

A system prompt is the standing instruction sent with every request, defining the model’s role, task, constraints and output format.

Its economics differ from everything else in a request. A user message is paid for once; a system prompt is paid for on every request, by every user, and in a chat product on every turn of every conversation.

That makes it both the highest-leverage text in an application and the most expensive place to leave something that changes nothing.

In practice

System prompts leak. Assume any instruction in one can be extracted and quoted back, and write it accordingly: no credentials, no internal hostnames, no wording you would not want a user to read. It is also the most cacheable part of your prompt, which means it belongs first and should not contain anything that changes per request.

Common questions

How long should a system prompt be?

As long as it needs and no longer — every request pays for every line. The useful discipline is not a word limit but a review: read it line by line each quarter and delete what stopped mattering.

Can users see the system prompt?

Assume yes. System prompts leak through extraction attempts, error messages and logs. Write anything you put in one as if it will be quoted back: no credentials, no internal hostnames, no wording you would not defend publicly.

Measure it

  • Prompt engineeringPrompt engineering is the practice of structuring the text sent to a language model so that its output is reliable enough to build on.
  • Cached inputCached input is prompt content the provider has already processed and retained, billed at roughly a tenth of the base input rate.
  • Few-shot promptingFew-shot prompting supplies a handful of worked examples in the prompt so the model infers the pattern rather than being told it.

More in Prompting